Supervised Transfer Learning Framework for Fault Diagnosis in Wind Turbines

1. Problem Statement

The paper addresses critical challenges in the condition monitoring and fault diagnosis of Wind Turbines (WTs):

• Data Scarcity & Labeling: Faults in WTs are rare events, leading to a lack of labeled data for training machine learning models.
• Domain Specificity: Traditional supervised learning approaches often require building a separate model for each individual turbine. This is inefficient because turbines may have different manufacturers, components, and signal patterns, making models non-transferable.
• Interpretability: Many existing transfer learning methods (e.g., deep neural networks) produce abstract, unexplainable features. Diagnosticians require clear reasoning behind a fault detection to trust the system and distinguish between actual mechanical faults and sensor errors.
• Sensor vs. Mechanical Faults: Distinguishing between a faulty sensor (cheap to replace) and a mechanical failure (e.g., bearing fault, which can cause total machine failure) is crucial to prevent unnecessary maintenance or missed critical failures.

2. Methodology

The authors propose a Supervised Transfer Learning Framework that operates within a specialized feature space called the "Anomaly-Space."

A. The Anomaly-Space (Feature Engineering)

Instead of feeding raw SCADA or vibration data directly into a classifier, the framework utilizes a pre-processed feature space provided by the research partner (EnBW). This space converts raw data into normalized anomaly scores for specific components.

• Input Data: SCADA data (temperature, pressure, etc.) and high-frequency vibration data.
• Detectors:
  • Tuplet Detector: Monitors groups of semantically similar components (e.g., generator temperatures across three phases). It calculates the variance of these groups. A significant increase in variance indicates a sensor fault.
  • Broad-Band-Characteristic-Value (bbcv) Detector: Processes vibration data using Fast-Fourier-Transformation (FFT) to extract features (skewness, kurtosis, mean). It then performs trend analysis (hypothesis testing) to detect increasing trends indicative of bearing faults.
• Normalization: Scores are normalized such that values > 1.0 are considered anomalous. This makes the features intuitive for human diagnosticians.

B. Data Preprocessing

• Labeling: Data within fault time frames (defined by repair dates) are labeled with the specific fault type; otherwise, they are labeled "Normal."
• Windowing: A sliding window approach (window size = 144 samples, stride = 1) is used to capture temporal relationships.
• Feature Extraction: From the Anomaly-Space, two new features are extracted per window:
  1. Trend-Certainty (tc): Derived from the Mann-Kendall test to determine if a trend exists.
  2. Variance (var): Statistical variance within the window.
• Final Feature Set: The framework reduces the input to 6 features (outputs from detectors + extracted trend/variance features).

C. Transfer Learning Strategy

The approach is classified as a Feature-Based Transfer Learning method.

• Mechanism: The Anomaly-Space acts as a "domain-shared feature space." By mapping data from different turbines into this space, the distribution discrepancy between source (training) and target (testing) domains is minimized without needing complex GANs or instance reweighting.
• Goal: Train a single classifier on data from multiple source turbines and apply it to new, unseen turbines (target domains).

D. Models Evaluated

The authors compared three popular supervised classifiers:

1. Random Forest (RF)
2. Light-Gradient-Boosting-Machines (LightGBM)
3. Multilayer Perceptron (MLP)

3. Experimental Setup

• Dataset: Data from 7 Wind Turbines across 5 different wind parks.
  • Train/Validation Set: 5 turbines from 4 parks (Cases 1–6). Includes bearing faults and sensor faults.
  • Test Set: 2 turbines from 2 parks (Cases 7–8). Crucially, one test turbine is from a completely different wind park than the training set, rigorously testing cross-domain capabilities.
• Fault Types: Bearing faults (Fast Shaft Bearing) and Sensor faults (Temperature sensors).
• Evaluation Metric: Fβ-score with β = 0.5. Precision is weighted more heavily than recall because false positives (misdiagnosing a healthy turbine as faulty) are costly and erode trust in the system.

4. Key Results

• Model Performance (Training):
  • The Multilayer Perceptron (MLP) outperformed RF and LightGBM.
  • MLP Configuration: 1 hidden layer with 5 neurons, ReLU activation, Adam optimizer, learning rate 0.001.
  • Cross-Validation Score (Train): F0.5 score of 0.874.
• Model Performance (Testing):
  • The MLP was trained on the full training set and evaluated on the unseen test set (including the turbine from a different park).
  • Test Score: F0.5 score of 0.937.
  • Precision: 0.992 (Very high, indicating few false alarms).
  • Recall: 0.789.
• Observation: The model performed better on the test set than the cross-validation training set. The authors attribute the lower training score to "loose contact" sensor faults in the training data, where signals intermittently appeared normal, confusing the model.

5. Key Contributions

1. Interpretable Feature Space: The proposal of an "Anomaly-Space" that translates complex raw sensor data into intuitive anomaly scores (>1.0), making the decision process transparent for diagnosticians.
2. Effective Cross-Domain Transfer: Demonstrated that a single classifier trained on data from 4 different wind parks can successfully diagnose faults in turbines from a completely different park, reducing the need for per-turbine models.
3. Comprehensive Evaluation: Conducted extensive comparisons of RF, LightGBM, and MLP using stratified cross-validation on real-world industrial data.
4. Practical Framework: Provided a complete end-to-end pipeline from raw SCADA/vibration data to fault diagnosis, specifically addressing the distinction between sensor and mechanical faults.

6. Significance and Future Work

• Significance: This framework offers a scalable solution for the wind energy sector. By using a single model for multiple turbines, it significantly reduces the computational and engineering overhead of maintaining individual models. The high precision ensures that maintenance teams trust the alerts, while the interpretability helps in root cause analysis.
• Future Work: The authors suggest integrating Out-Of-Distribution (OOD) detection to identify fault types that the model has never seen before, further enhancing the robustness of the diagnostic system.

In conclusion, the paper successfully validates that a feature-based transfer learning approach using an interpretable Anomaly-Space is a highly effective strategy for automated, cross-domain fault diagnosis in wind turbines.